Research On Rectification Techniques Of Facade Images

Image-based architecture modeling is a famous application in both computervision and computer graphics. Image-based modeling obtains the three-dimensionalinformation of the target from single or more images, and then utilizes the informationto reconstruct the photo-realistic three-dimensional target. In fact the rectified imageis usually used as the input of image-based modeling method, hence imagerectification has a great significance for image-based modeling and even computervision. In this work, we focus on rectification techniques of the fa ade image.Generally, the fa ade image contains rich geometric structures, such as asymmetric structure or a repeating structure, hence this regular image usually has thelow-rank property. The original TILT considers that the rank of rectified fa ade imageis lower than the original input image. However, in practice, the fa ade image is oftencorrupted by shadows, which may be caused by neighboring high buildings or someself-protruding parts. In such situations, the rank of rectified fa ade image may not belower as compared with the original input image. Hence, we propose an improvedmethod to rectify fa ade images which are corrupted by shadows. The new method isdeduced from the original TILT, and improved by introducing a multiplicative shadowfactor. In our model, the rectified fa ade image is decomposed into a shadow-freelow-rank image, a multiplicative shadow image and the additive noise corruption, i.e.,the model is that the rectified image equals to the low-rank image multiplied by theshadow image, and yet with the additive noise corruption. Meanwhile, in this model，the shadow image is supposed to be smooth and the noise image be sparse. Finally, inthis paper, we do experiments on both synthetic images and real fa ade images.Experimental results demonstrate that our method provides more accurate andstable rectification results as compared with the original TILT, especially whenshadows are strong in the input images.